Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective

Xueying Bai, Yifan Sun, Niranjan Balasubramanian

Research Area: Learning algorithms for LMs

Keywords: Continual Learning, Attention Sink, Interference, Over-Smoothing, Pre-Scaling

TL;DR: We tackle the LM's interference in CL from an attention sink's perspective, where different models' concentration on common tokens (e.g., special tokens and punctuations) may influence their CL performance besides the models' single-task capacities.

Abstract: Continual learning (CL) aims to train models that can sequentially learn new tasks without forgetting previous tasks' knowledge. Although previous works observed that pre-training can benefit CL, it remains unclear whether a pre-trained model with higher downstream capacity also performs better in CL. In this paper, we observe that pre-trained models may allocate high attention scores to some 'sink' tokens, such as [SEP] tokens, which are ubiquitous across various tasks. Such attention sinks may lead to models' over-smoothing in single-task learning and interference in sequential tasks’ learning, which may compromise the models' CL performance despite their high pre-trained capabilities. To reduce these effects, we propose a pre-scaling mechanism that encourages attention diversity across all tokens. Specifically, it first scales the task's attention to the non-sink tokens in a probing stage, and then fine-tunes the model with scaling. Experiments show that pre-scaling yields substantial improvements in CL without experience replay, or progressively storing parameters from previous tasks.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the COLM Code of Ethics on https://colmweb.org/CoE.html

Author Guide: I certify that this submission complies with the submission instructions as described on https://colmweb.org/AuthorGuide.html

Submission Number: 431